Optimal Transport Theory for Machine Learning with Limited and Less Labels
Abstract
Due to the exponential growth of unlabeled data collected and the hardness, labor-expensiveness, errors, and time-consuming of the process to label data, learning with less and limited labels (LwLL) has emerged as an important research endeavor in machine learning and deep learning. LwLL aims to escalate learning process with limited labeled data similar to the way human learns a new concept effortlessly. It covers a set of related problems, notably transfer learning, domain adaptation (DA), meta learning, domain generalization, and semi-supervised learning. In parallel, optimal transport (OT) is a recent powerful mathematical theory that has been rapidly become a mainstream research tool in machine learning. With its attractive geometry interpretation, computational tractability and expressiveness, OT offers a principal tool to address several LwLL problems. However, this connection is still very limited and under-explored in the current literature. To this end, this proposal aims to investigate OT theory for LwLL- a problem which, to our best knowledge, is new and novel. In particular, we plan to discover a bridging theory connecting OT and LwLL that can elegantly exploit the unique characteristics of OT transport (e.g., distribution matching and clustering view) for advancing the current state-of-the-art in LwLL.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA23862114049XX0
Entities
People
- Trung Le
Organizations
- Air Force Office of Scientific Research
- Monash University
- United States Air Force